This paper proposes the use of Extreme Learning Machine regressionframework to generate virtual frontal view from its corresponding side view. Kernel version of ELM is used for the non-linear mappingestimation between frontal and its corresponding non frontal view. Non Linearestimation using proposed KELM is also found to be satisfactory even in the caseof compressed images. In addition to regression framework, this paper alsoexplored few popular face features like LBP, HOG, appearance and entropyfeatures, with ELM classification framework.
Combination of appearance andentropy feature set is found to give best classification accuracy among theconsidered feature set. Proposed experiments are conducted on subset of FERETdatabase with images upto 45° pose variation. Keywords: ELM(ExtremeLearning Machine),KELM(Kernel Extreme Learning Machine).1 IntroductionDuring last fewyears, research for accurate and reliable face recognition has madeconsiderable progress.Machine vision face recognition depends largely on features extracted from eye,nose and mouth regions of the input face. Thesediscriminant features becomes occluded in case of face pose like profile pose.Recognition capability degrades further due to variability in illumination andscales.
To achieve the objective of robust face recognition across poses,various approaches like: pose normalization, pose adaptive filters and posesynthesis has been proposed in the literature. All these approaches extracthigh dimensional features (local/global) from the images and integrate them byusing learning algorithms. Extreme Learning Machine (ELM) proposed by Huanget-al. 1,2 for single hidden layer feed- forward networks, has gainedimportance among various learning algorithms.
Speed advantage and convex modelstructure of ELM is responsible for its usage across various domains. ELM mapsthe input features to the feature space using non-linear activation function1-3. In the mapping process of the input space RD to a highdimensional feature space RL (ELM Space), the dimensionality L of the ELM space is usuallyempirically chosen. In order to avoid the application of time-consumingalgorithms for the determination of the ELM space dimensionality, kernelversions of the ELM classifier have been proposed (3, 6). The idea in kernel versions of the ELM 3,6is to avoid the direct calculation of the network hidden layer outputs, and theinclusion of the inherent encoding in the so-called ELM kernel matrix definedby K = ?T?, where ? ? RL×N refers to the training data representationsin the ELM space and N is the number of training data. Thisapproach however contradicted the standard definition of kernel matrix.
Parviainen et al 8 used Cholesky decomposition of the kernel matrix definedon the input training data, in order to calculate an appropriate matrix ?? RN×N. In 9 showed that ELM kerneldefinition can be adopted for the calculation of the ELM kernel matrix for theRBF and sigmoid hidden layer activation functions. For appropriate ELM spacedetermination they had used a low-rank decomposition of the kernel matrixdefined on the input training data. Recently, Goel et al.
9 used kernel ELMfor pose synthesis.In thispaper we explored both regression and classification framework of ELM torecognize faces across the poses. Non-linear regression via kernel ELM isproposed to estimate the non-linear mapping between frontal face views from itscounter-part non-frontal views for effective face recognition. This work avoidstime consuming step for finding optimal value for hidden neurons byexperimental tuning. Compared to linear regression, KELM regression is moreefficient for virtual view generation as it considers non-linear shape of theface view.
Various features like LBP, HOG, entropy and intensity features areexplored with ELM classification framework. The rest of this paper is organizedas follows: In Section 2, we give an outline of ELM classification framework aswell as ELM regression framework for pose normalization. Section 3 presents theoutline of the proposed pose normalization approach. In Section 4 resultsobtained with FERET face pose images upto 45 degree deviation are analyzed andinterpreted. Finally, in section 5 we conclude the paper and highlight some ofthe future ideas for research.